App Store Review Arbitrage Agent Skill

app-store-review-arbitrage

Convert a competitor's App Store or Google Play URL into a one-session GTM brief: ranked complaint clusters, a broken promise map, landing page headlines, and ad copy directions — all sourced from verbatim reviews.

Critical Rules (read before Step 1)

These rules apply throughout all steps. Violating any of them fails Self-QA (Step 6).

• Every quote must be verbatim. No paraphrase, no grammar correction, no cleaning. Exact reviewer words only.
• No fabricated statistics. Do not write "40% faster" or "2× more reliable" unless a reviewer explicitly used similar language. The Self-QA step checks for uncited percentages.
• Cluster names must use reviewer language. Study the anti-pattern table in Step 3.
• Every headline and ad copy direction must cite its source cluster. Format: [cluster: "cluster-name"].
• Section 2 is always present in the output — even when degraded. Never skip or omit it.
• No banned words in any generated copy: powerful, robust, seamless, innovative, game-changing, streamline, leverage, revolutionize, transform.

Step 1 — Parse Input and Detect Platform

Accept a natural language prompt containing one app URL. Extract the URL.

Platform detection:

• apps.apple.com → App Store
• play.google.com/store/apps/details?id= → Google Play
• Any other URL → stop and respond: "Please provide a direct App Store or Google Play URL. I can't analyse review data from other sources."

ID extraction (do this before calling the script):

| Platform | What to extract | How |

|---|---|---|

| App Store | Numeric app_id | Digits after /id in the URL |

| App Store | country | 2-letter code after apps.apple.com/ (e.g., us, gb) |

| Google Play | package_name | Value of id= query parameter |

Persist the extracted values — you will need them for the output filename in Step 7.

If product_context was provided in the user's prompt (what their own product does), store it — used to personalise copy in Step 5.

Step 2 — Collect Reviews & Metadata

Run the full fetch script:

python3 scripts/fetch_reviews.py "{app_url}" --output {tmpdir}/asr-raw.json

*(Note: Replace {tmpdir} with your operating system's temp directory, e.g., /tmp on macOS/Linux or C:\Temp on Windows).*

This fetches both the store description metadata and the reviews.

• App Store: iTunes API — free, no auth. App Store reviews are fetched via Apple's public iTunes RSS feed. Some apps return 0 reviews due to Apple's API limitations — in that case the skill continues with available data and logs a warning. Google Play is the primary supported path.
• Google Play: google-play-scraper package — free, no auth

If the script fails, read the error from stderr. Common causes:

• Package not installed: run pip install google-play-scraper
• App not found: verify the URL is a current, live listing
• Google Play API error: run pip install --upgrade google-play-scraper and retry

The script will print collection progress to stderr. Wait for it to complete. After completion, read {tmpdir}/asr-raw.json and display the collection summary to the user:

✓ Collected [N] reviews ([N] low-star 1–3★) from [platform]
  Date range: [oldest] to [newest]
  Package: [iTunes API | google-play-scraper]

Check the exit code:

• Exit 0 → collection succeeded, check metadata.store_description. If null: note this — Section 2 will use the degraded state. Proceed to Step 3.
• Exit 1 → error (read stderr message, surface it to user, stop)
• Exit 2 → Gate 1 triggered (< 10 low-star reviews found)

Gate 1 — Low signal stop: If the script exits with code 2, read the gate_message from {tmpdir}/asr-raw.json and surface it to the user verbatim. Do not proceed to Step 3. Do not produce a partial brief.

Step 3 — Complaint Clustering

Load low_star_reviews from {tmpdir}/asr-raw.json.

Cluster all low-star reviews into 4–6 named complaint themes. Apply this formula to score each review:

complaint_weight = (4 - rating) × recency_factor

recency_factor:
  review age ≤ 90 days  → 1.0
  review age 91–365 days → 0.7
  review age > 365 days  → 0.4

review age = (today's date) − (review date field) in days.

cluster_score = sum of complaint_weight for all reviews in the cluster.

Cluster naming — critical rule:

You will want to write abstract names. Resist. Use the exact verb and noun from reviews.

| ❌ Abstracted (wrong) | ✅ Reviewer language (correct) |

|---|---|

| "Stability issues" | "Crashes when exporting to PDF" |

| "Sync problems" | "Data lost after sync between phone and desktop" |

| "Monetisation friction" | "Paywall appears after 3 days, not 14 as promised" |

| "Performance degradation" | "App freezes every time I search" |

| "Onboarding issues" | "Can't figure out how to invite a teammate" |

Rules:

• Each review belongs to exactly one cluster (assign to its dominant theme)
• Discard any cluster with fewer than 3 reviews — log it as noise
• Select 3–4 verbatim quotes per cluster: lowest star rating first, then most recent

Gate 2 — Minimum cluster size: After discarding sub-3-review clusters, check how many clusters remain.

Gate 3 — Low-confidence flag: If fewer than 3 clusters remain:

• Do NOT stop. Continue to output.
• Prepend this to the brief header immediately after the app metadata:

> ⚠ LOW CONFIDENCE: Only [N] complaint cluster(s) met the minimum evidence threshold (≥ 3 supporting reviews). Output reflects limited data. Consider a competitor with more reviews, or broaden the rating filter.

• Include Medium-tier clusters in the output (score ≥ 5)

Tier classification (for the leaderboard table in Section 1):

• Critical: score ≥ 60
• High: score 15–59
• Medium: score 5–14 (include only when Gate 3 applies)
• Noise: score < 5 (discard, do not include)

Write clusters to {tmpdir}/asr-clusters.json:

{
  "clusters": [
    {
      "name": "cluster name in reviewer language",
      "score": 34.5,
      "tier": "High",
      "review_count": 14,
      "verbatim_quotes": [
        {"rating": 1, "text": "exact reviewer words", "date": "YYYY-MM-DD"},
        ...
      ]
    }
  ],
  "discarded_noise": 2,
  "gate_3_triggered": false
}

Step 4 — Broken Promise Detection

This is the step that differentiates this skill from every existing tool. It must run as a distinct, named step.

Load:

• metadata.store_description from {tmpdir}/asr-raw.json
• All clusters from {tmpdir}/asr-clusters.json

If store_description is null: Set store_description_available: false. Write {tmpdir}/asr-promises.json with empty broken_promises array and detection_note as specified below. Proceed to Step 5.

If store description is available:

• Extract claims. A claim is any specific, testable assertion about app behavior. See references/broken-promise.md for the full definition and examples. Exclude vague superlatives, team descriptions, and press quotes.

• Cross-reference. For each claim, check all cluster names and verbatim quotes. A contradiction exists when the cluster directly documents failure of the promised behavior (minimum 3 reviews).

• Produce broken promise records — one per confirmed contradiction:

   {
     "claim_text": "verbatim excerpt from store description",
     "complaint_cluster": "exact cluster name",
     "gap_label": "Claims X; users report Y",
     "evidence_count": 18
   }

Write to {tmpdir}/asr-promises.json:

{
  "store_description_available": true,
  "broken_promises": [...],
  "no_contradictions_found": false,
  "detection_note": null
}

Degraded states:

• No description: store_description_available: false, detection_note: "Store description unavailable (fetched YYYY-MM-DD, returned empty). Broken promise comparison cannot be performed."
• No contradictions: no_contradictions_found: true, detection_note: "No broken promises detected. Store description does not appear to overclaim relative to complaint clusters."

See references/broken-promise.md for anti-patterns (what NOT to flag).

Step 5 — Generate Copy

Using clusters from Step 3 and broken promises from Step 4, generate Sections 3–5 of the brief.

Copy rules (apply to all three sections):

• Every headline and direction must cite its source cluster: [cluster: "cluster-name"]
• No banned words: powerful, robust, seamless, innovative, game-changing, streamline, leverage, revolutionize, transform
• No fabricated statistics — no percentages or numbers unless a reviewer used them
• If product_context was provided: make "Say this" directions specific to that product's features. If not: write as positioning templates the user fills in.
• Use reviewer language in headlines — derive from or quote actual review text

Section 3 — Landing Page H1 Bank (3–5 headlines):

• Each: "[headline text]" [cluster: "cluster-name"]
• ≤ 8 words where possible
• Address the frustrated user directly

Section 4 — Ad Copy Directions (exactly 3 pairs):

Cluster: [cluster name]
Not that: "[what the competitor claims or a generic weak alternative]"
Say this: "[counter-claim grounded in complaint evidence]"
Evidence: [N] reviewers reported [verbatim complaint summary]

Section 5 — Anti-Claim Warnings:

• One warning per broken promise from Step 4
• If Section 2 is degraded: single note (see references/brief-format.md for exact wording)

Step 6 — Self-QA

Before saving, verify the generated brief against these checks. If any check fails, fix the specific item and re-verify — do not save a failing brief.

| Check | Rule |

|---|---|

| Verbatim quotes present | Every cluster has ≥ 2 verbatim quotes |

| No banned words | None of the 9 banned words appear in Sections 3–5 |

| No uncited percentages | Any % in output must trace to a reviewer's actual words |

| All copy cited | Every headline and "Say this" has a [cluster: "name"] citation |

| Cluster count ≥ 1 | At least one cluster survived Gates 2/3 |

| Section 2 present | Section 2 appears in the output (in any state) |

| Quote ratings ≤ 3 | All verbatim quotes came from 1–3★ reviews |

Note on cluster count: The minimum for a passing brief is 1 cluster (not 3). The Gate 3 low-confidence flag handles cases where < 3 clusters survive — that is a warning, not a failure. Self-QA fails only if 0 clusters exist.

Step 7 — Save Output

Assemble the full brief per the format in references/brief-format.md.

Save to:

docs/review-briefs/[app-id]-[YYYY-MM-DD].md

Create the docs/review-briefs/ directory if it does not exist.

Print the full brief to the user.

Clean up temp files: {tmpdir}/asr-raw.json, {tmpdir}/asr-clusters.json, {tmpdir}/asr-promises.json.